AAU Student Projects - visit Aalborg University's student projects portal
A master's thesis from Aalborg University
Book cover


Bringing Reality to Virtual Worlds: An Evaluation of 3D Gaussian Splatting for Accessible 3D Reconstruction in Virtual Production

Author

Term

4. term

Education

Publication year

2024

Pages

9

Abstract

Denne afhandling undersøger, om 3D Gaussian Splatting (3DGS) kan fungere som et tilgængeligt og effektivt alternativ til fotogrammetri i virtuel produktion (VP). Med fokus på mid-range optageudstyr og let tilgængelige værktøjer blev rekvisitter og steder med stigende kompleksitet scannet med smartphone, iPad (LiDAR/Polycam) og en lille drone. Data blev behandlet både i en lokal 3DGS-implementering (Jawset Postshot) og på en cloud-platform (Luma AI), hvorefter de genererede 3D-modeller blev vurderet for geometrisk nøjagtighed, teksturkvalitet og realtidsrendering samt praktiske forhold som behandlingstid, ressourceforbrug og integration i VP-workflows. Foreløbige resultater peger på, at både Postshot og Luma AI er lovende alternativer til fotogrammetri: Luma AI udmærker sig ved hastighed og brugervenlighed, mens Postshot tilbyder større kontrol og tilpasning, med resultater der afhænger af objektets kompleksitet og optagemetode. Studiet dokumenterer tekniske udfordringer i tilgængelige pipelines og fremhæver 3DGS’ potentiale til at demokratisere 3D-rekonstruktion i VP. Fremtidigt arbejde vil udvide evalueringen til Neural Radiance Fields (NeRFs), særligt for generering af nye synsvinkler.

This thesis examines whether 3D Gaussian Splatting (3DGS) can serve as an accessible and efficient alternative to photogrammetry in virtual production (VP). Using mid-range capture devices and readily available tools, props and locations of increasing complexity were scanned with a smartphone, iPad (LiDAR/Polycam), and a small drone. The datasets were processed with a local 3DGS implementation (Jawset Postshot) and a cloud platform (Luma AI), and the resulting 3D models were evaluated for geometric accuracy, texture fidelity, and real-time rendering, alongside practical factors such as processing time, resource demands, and integration into VP workflows. Preliminary findings indicate that both Postshot and Luma AI offer promising alternatives to photogrammetry: Luma AI emphasizes speed and ease of use, while Postshot provides greater control and customization, with outcomes influenced by asset complexity and capture method. The study documents technical challenges in accessible pipelines and highlights the potential of 3DGS to democratize 3D reconstruction in VP. Future work will extend the evaluation to Neural Radiance Fields (NeRFs), particularly for novel view synthesis.

[This summary has been generated with the help of AI directly from the project (PDF)]